AMD 2008 thoughts?

[aussiestilgar]

AMD will be back in the same position they were before K7... as the underdog. AMD surprised Intel once, now it seems Intel are on top of their game, so to speak. The question is whether AMD can surprise Intel again. If so, methinks it won't happen until 2009 or after.

 

[jones377]

Originally posted by: Martimus

Originally posted by: Kuzi

Originally posted by: Idontcare
And there are no DEC alpha design teams left for them to hire-in to develop another K7.

I suppose they could snag a SUN or IBM design team and bring them in to inject some fresh ideas into their internal stalled out design team...

I mean come on, look at that THG article comparing Stars core versus K8...that's all they got for 4 years of investment? THAT is what AMD management was holding out for instead of doing an MCM with K8 to release a quad chip in the meantime? Yikes!

I have to disagree here. I think AMD has some of the brightest design teams in the world. Their problems seem more related to bad decisions made than anything else, and probably the fact that they have a lot less cash than Intel.

Had AMD set out to just improve on single core performance when compared to K8 some years back, I'm sure the end result would be better than a 15% (average) improvement. But maybe the native quad core design took more time and resources than they had expected, and that cost them too much in the end.

Imagine a dual core K10 that performs 30% better than K8, then AMD just sticks two of those together on the same die like Intel does and makes a quad core. That would have worked better for them performance wise, manufacturing wise, and probably would have been released a while ago.

AMD is just very late, their 65nm process is messed up, and now they need a great 45nm launch as soon as possible or it won't be much of a fight.

You forget that sticking two dual core processors with seperate IMC's on a single die is a difficult thing to get working. You may need to have each one control a seperate DIMM slot (Like Quad FX does), which means you need to use at least 2 sticks of RAM, and to get the best performance 4 (Since it is dual channel, and AMD split up the IMC a while ago to increase bandwidth.) It may have been just as difficult to design a K8 MCM quad core as it was to creat a native quad core, and they figured that the native solution was a better solution anyway. Intel never needed to worry about that, because it uses a Memory Controller based on the North Bridge instead of the processor.

Or you disable the IMC on one of the cores and route a HT connection between them on the MCM. This has already been done on the board level for a few low-end dual socket Opteron motherboards in the past.

 

[Idontcare]

Originally posted by: jones377

Originally posted by: Martimus

Originally posted by: Kuzi

Originally posted by: Idontcare
And there are no DEC alpha design teams left for them to hire-in to develop another K7.

I suppose they could snag a SUN or IBM design team and bring them in to inject some fresh ideas into their internal stalled out design team...

I mean come on, look at that THG article comparing Stars core versus K8...that's all they got for 4 years of investment? THAT is what AMD management was holding out for instead of doing an MCM with K8 to release a quad chip in the meantime? Yikes!

I have to disagree here. I think AMD has some of the brightest design teams in the world. Their problems seem more related to bad decisions made than anything else, and probably the fact that they have a lot less cash than Intel.

Had AMD set out to just improve on single core performance when compared to K8 some years back, I'm sure the end result would be better than a 15% (average) improvement. But maybe the native quad core design took more time and resources than they had expected, and that cost them too much in the end.

Imagine a dual core K10 that performs 30% better than K8, then AMD just sticks two of those together on the same die like Intel does and makes a quad core. That would have worked better for them performance wise, manufacturing wise, and probably would have been released a while ago.

AMD is just very late, their 65nm process is messed up, and now they need a great 45nm launch as soon as possible or it won't be much of a fight.

You forget that sticking two dual core processors with seperate IMC's on a single die is a difficult thing to get working. You may need to have each one control a seperate DIMM slot (Like Quad FX does), which means you need to use at least 2 sticks of RAM, and to get the best performance 4 (Since it is dual channel, and AMD split up the IMC a while ago to increase bandwidth.) It may have been just as difficult to design a K8 MCM quad core as it was to creat a native quad core, and they figured that the native solution was a better solution anyway. Intel never needed to worry about that, because it uses a Memory Controller based on the North Bridge instead of the processor.

Or you disable the IMC on one of the cores and route a HT connection between them on the MCM. This has already been done on the board level for a few low-end dual socket Opteron motherboards in the past.

That is the primary point of logic I am drawn to everytime I read a post wherein the poster is arguing that MCM for AMD's K8s is nigh impossible...dual-socket mobo makers already proved it could be done though...dual-socket mobos can be viewed as elaborate macro-scale implementations of MCM...and they work.

Not arguing that it is economical, am just saying the proof that it can be done has already been delivered to the market...it was AMD that decided to not take it to its next logical step (yet).

 

[Rottie]

I really hate Intel so much but I am force to give up AMD for Core 2 Duo on my next build in 2 years.
I want to thank AMD for outstanding A64 X2 4800+ cpu but all of my friends with Core 2 Duo are laughing at me for being so SLOW cpu I feel that AMD for 2008 outlook is not going to be FASTER cpu anymore. Intel is going to stay superfast for many years to come. AMD doesn't have have enough money to make faster cpu to kill Intel cpu for sure.

Now I must say...Good Bye AMD!

 

[Martimus]

I didn't say that it was impossible, just that it was difficult. They needed to make a design change either way, and they likely felt that a native solution would be the better solution. Of course, the poor yields on a native solution vs. a MCM solution may overcome the benefits of the native solution. I can see why the wanted a native quad core, because seeing how they perform compaired to faster Intel chips on Anandtechs Server tests show just how much better that is on multithreaded applications. Of course, there is only a small market share that uses those types of high bandwidth multithreaded applications.

 

[BitByBit]

I very much doubt K10's superior thread scaling is due to its single-die design. The X2 was predicted to scale much more linearly than the Pentium D in multithreaded apps, but it didn't; if memory serves, they were roughly equivalent. K10's unganged memory controllers and memory access prediction algorithms are a far more likely explanation for K10's thread scaling.

It's been stated before: AMD's monolithic approach to multicore architecture has cost them more than it's worth. Integrating the memory controller was a good move, but AMD's obsession with single-die design is undoubtedly holding them back. It seems raw performance doesn't care about elegance.

 

[Martimus]

I wouldn't call it an obsession. They made a design decision. It may have been the wrong one, but they can't afford to go ahead with both ideas, so they chose one. That is it. They may go with an MCM design for their next upgrade to 8 cores, but the ship is already sailed on this one, and it is likely cheaper and easier to fix the design they have than go back and make a 4 core MCM K8 chip.

 

[Kuzi]

AMD already passed that hurdle in designing a native quad core, and K10 is running but maybe not as well as they had hoped. So you are right in saying they should work on fixing it now and maybe tweaking the performance if possible.

About the the superior multithread scaling on K10, I would say the IMC also has something to do with it. And Nehalem will have an integrated MC so any advantage K10 might still hold in efficiency or in servers will probably be gone when it is released.